Where Should We Start With AI Agents?

Short answer

Start with one recurring business workflow that has a clear owner, measurable friction, accessible information, and manageable consequences if something goes wrong. Then ask whether it needs an agent that chooses its next steps, or whether simpler automation would work. Test the smallest suitable solution with limited permissions and human oversight before expanding its responsibilities.

The starting pattern

  1. 01One team
  2. 02One workflow
  3. 03One measurable outcome

A useful starting pattern is one team, one workflow, and one measurable outcome.

Start with the problem

Why start with a business problem?

“Deploy an AI agent” describes a technology decision. “Reduce the time required to prepare support cases for review” describes an operational objective.

The second gives your team something to evaluate: the current process, its exceptions, the systems involved, and the result that would justify investment.

Before discussing platforms, document:

  • What triggers the work and what counts as completion.
  • Who performs it and who owns the outcome.
  • Where time, errors, delays, or handoffs accumulate.
  • How frequently it occurs.
  • What happens when the process fails.

Choose a workflow that matters enough to measure but is narrow enough to understand. An unclear process needs clarification before it needs autonomy.

Agent or automation?

Do you actually need an AI agent?

Complexity ladder

  1. 01Deterministic automation
  2. 02Retrieval-based assistant or chatbot
  3. 03Copilot
  4. 04AI-assisted workflow
  5. 05Bounded AI agent
Each step adds autonomy and operating work. Climb only as far as the work requires.

An AI agent uses a model to choose actions and tools as it works toward a goal. An AI-assisted workflow can use the same underlying models while following steps defined in software.

This distinction matters: Anthropic recommends increasing complexity only when needed, recognizing that simpler approaches can be sufficient. Building effective agents

Scroll horizontally to compare all columns.

Do you actually need an AI agent?
What the work requiresConsider firstExample
Stable rules and predictable inputsDeterministic automationRoute a request using a required department field
Answers from approved informationRetrieval-based assistant or chatbotFind the current travel policy
Drafting while a person directs the workCopilotPrepare a response for an employee to edit
Language understanding within fixed stepsAI-assisted workflowExtract document fields, validate them, then request approval
Choosing different steps and tools as circumstances changeBounded AI agentInvestigate a support request across approved systems and propose a resolution

These categories can overlap. “Chatbot” describes an interface; it does not establish what happens behind it.

Use an agent when adaptive execution adds value. Summarizing text, extracting fields, or answering questions alone does not establish that need.

Choose a use case

How to identify your first AI agent use case

Starting point framework

  1. 01Candidate workflows
  2. 02Define the outcome
  3. 03Check agentic need

    Same sequence every time → fixed workflow first

  4. 04Apply risk gates

    Any gate fails → pause the candidate

  5. 05Score candidates
  6. 06Pilot candidate

Use the AI Agent Starting Point Framework to compare candidates before committing to a build.

  1. Define the outcome

    Write one sentence:

    Help the support team prepare account-access cases for review by gathering relevant records and recommending the next step.

    Then establish baseline handling time, quality, volume, and exception rates.

  2. Check whether agentic behavior is necessary

    Ask: Does the system need to decide what to investigate next based on what it finds?

    If every case follows the same sequence, test a fixed workflow first. If the main need is better information access, start with retrieval.

  3. Apply readiness and risk gates

    Pause a candidate if:

    • No one owns the process or can judge success.
    • Required information cannot be accessed appropriately.
    • Permissions cannot be restricted to the task.
    • Material errors cannot be detected, contained, or reviewed.
    • The team cannot support the workflow after launch.

    A strong business case does not override these conditions.

  4. Score the remaining candidates

    Rate each factor from 1 to 5, with 5 always meaning more favorable for an initial pilot.

    Scroll horizontally to compare all columns.

    4. Score the remaining candidates
    FactorWhat a score of 5 means
    Business impactSolves a meaningful operational problem
    Technical feasibilityRequired behavior is achievable with available capabilities
    Data readinessRelevant information is accessible, current, and usable
    Integration simplicityFew, well-supported connections are required
    Risk manageabilityConsequences are limited and controls are practical
    Time-saving potentialSubstantial repetitive effort could be removed
    Implementation effortA narrow scope can be delivered with available capacity
    MeasurabilitySuccess can be judged against a clear baseline

    Add the scores for a total out of 40. This equal-weight worksheet helps structure discussion; it does not predict ROI. Business impact covers the broader outcome, while time-saving potential isolates labor effort.

  5. Worked example: three possible starting points

    Illustrative scores, not research findings or customer results. Assume all three candidates have passed the initial gates.

    Scroll horizontally to compare all columns.

    Worked example: three possible starting points
    FactorInternal policy retrievalSupport triage and investigationInvoice exception preparation
    Business impact345
    Technical feasibility543
    Data readiness443
    Integration simplicity542
    Risk manageability442
    Time-saving potential344
    Implementation effort542
    Measurability555
    Total34/4033/4026/40

    Policy retrieval may be the best first AI project, yet a retrieval assistant may be sufficient. Support investigation is the stronger agent candidate if each case requires different information-gathering steps. Invoice preparation deserves further investigation but has more integration and control dependencies.

    Change the scores using your own evidence. Treat close totals as a reason to inspect assumptions, not as a definitive ranking.

Worked example totals

  • Internal policy retrieval34/40
  • Support triage and investigation33/40
  • Invoice exception preparation26/40
Illustrative scores out of 40, not a definitive ranking.

Promising workflows

Which business processes are promising starting points?

The following are candidate patterns, not universal recommendations.

  • Internal knowledge and employee assistance

    Problem

    Employees repeatedly search policies and operating guidance.

    Pilot

    Search approved documents and return answers with source links. Add agentic investigation only if questions require adaptive searches across systems.

    Needs

    Maintained documents, document owners, and access controls.

    Measure

    Answer usefulness, factual support, and time to find information.

    Main risk and oversight

    Outdated or unauthorized information. Preserve user permissions and let employees flag answers for review. Often suitable as a first AI project without an agent.

  • Customer support triage and IT service desk assistance

    Problem

    Staff gather context before they can act.

    Pilot

    Read a ticket, consult approved records, choose relevant diagnostic checks, and draft a recommendation.

    Needs

    Ticketing data, knowledge articles, and scoped account or asset access.

    Measure

    Preparation time, routing accuracy, rework, and successful escalation.

    Main risk and oversight

    Incorrect advice or unintended changes. Have staff approve customer messages, account changes, and access requests. A good agent candidate when investigation paths vary.

  • Document intake and finance exception preparation

    Problem

    Staff reconcile inconsistent submissions and chase missing information.

    Pilot

    Classify documents, extract fields, run coded validations, gather supporting records, and prepare an exception summary.

    Needs

    Approved storage, document samples, business rules, and relevant records.

    Measure

    Processing time, extraction quality, and review effort.

    Main risk and oversight

    Incorrect values reaching financial systems. Keep calculations in deterministic software and require approval before posting or payment. Fixed extraction pipelines may be enough; finance variants need tighter controls.

  • Sales operations, research, and workflow coordination

    Problem

    Teams repeatedly assemble briefings, prepare CRM updates, or investigate stalled handoffs.

    Pilot

    Gather approved information, identify gaps, and draft updates or next actions.

    Needs

    CRM or task records, approved sources, and named owners.

    Measure

    Accepted drafts, preparation time, corrections, and handoff delays.

    Main risk and oversight

    Unsupported claims, incorrect records, or unwanted outreach. Require review before external publication or material updates. Suitable when the information-gathering path varies; fixed reminders need ordinary automation.

Check readiness

What does your organization need before building?

You need enough readiness for the selected workflow, not a complete enterprise AI transformation.

Confirm:

  • Business ownershipA process owner, sponsor, and participating users.
  • Usable informationRepresentative examples, current source material, and known gaps.
  • Integration accessApproved APIs or connectors, test environments, and system owners.
  • PermissionsRestricted identities, user access boundaries, and controlled credentials.
  • GovernanceDecisions about permitted data use, retention, providers, and relevant obligations.
  • Human oversightClear approval points, escalation routes, and backup procedures.
  • Operational ownershipPeople responsible for monitoring, incidents, changes, and budget.

Microsoft’s adoption model similarly treats organizational capabilities alongside technology and data. Agentic AI adoption maturity model

Connect systems with controlled interfaces

Prefer supported APIs and connectors. Specify what each tool can read or change, validate its inputs, and record its actions. Design retries so they do not accidentally create duplicate records or transactions.

Enforce authorization in the connected systems or service layer. A prompt telling an agent to respect permissions is insufficient. Google Cloud’s agent identity guidance illustrates the use of distinct agent identities and restricted access. Agent Identity overview

Treat retrieved documents and messages as potentially untrusted input. Test whether misleading content can cause inappropriate tool use or data disclosure.

Start with one agent when an agent is justified

A single agent with a small toolset is a sensible initial architecture to evaluate. Multiple agents can separate specialized responsibilities, but introduce additional coordination, debugging, and operating work.

Add them when testing reveals a specific limitation that separation resolves. OpenAI’s guide discusses both single-agent and multi-agent approaches. Practical guide to building agents

Plan a pilot

How do you launch your first AI agent pilot?

A pilot should produce an investment decision supported by evidence.

Scroll horizontally to compare all columns.

How do you launch your first AI agent pilot?
PhaseMain workEvidence needed to proceed
Discover and prioritizeMap the workflow, baseline performance, and select a candidateAgreed problem, owner, scope, and success criteria
Assess readinessInspect data, integrations, permissions, and constraintsFeasible access and manageable risks
Design and buildDefine tools, boundaries, approvals, and fallback behaviorWorking bounded solution and representative tests
Test and measureEvaluate outcomes, actions, costs, failures, and user experienceResults against pre-agreed acceptance criteria
Deploy and operateIntroduce limited live use, monitoring, support, and change controlsNamed owners and evidence supporting continued operation

A 30/60/90-day plan can be useful as an illustration:

  1. Days 1–30

    Establish the baseline, assess readiness, and define the pilot.

  2. Days 31–60

    Build and test in a controlled environment.

  3. Days 61–90

    Run limited live use if readiness permits, then decide whether to expand.

Data remediation, procurement, integration access, and organizational approvals can extend any phase. Day 90 is not a production guarantee.

Begin with historical cases or shadow mode, where the system proposes actions without executing them. Introduce approved actions only after evaluation.

Your pilot brief should include exclusions, a test set, cost limits, acceptance thresholds, stop conditions, deliverables, and operating ownership. Avoid expanding the scope whenever the demo reveals another possible use.

Measure value

How should you measure success and ROI?

Measure the complete workflow, including review, correction, and recovery.

Scroll horizontally to compare all columns.

How should you measure success and ROI?
MeasureWhat to examine
Task successWhether the case reached an acceptable outcome
Output qualityFactual support, completeness, and usefulness
Action qualityCorrect tools, arguments, permissions, and approvals
Human effortReview, corrections, escalations, and exception handling
Operational performanceCompletion time, failures, retries, and availability
EconomicsCost per successfully completed case
AdoptionWhether intended users continue using the solution

Google’s evaluation documentation distinguishes final-response assessment from tool-trajectory assessment. A plausible answer does not establish that execution was correct. Evaluate Gen AI agents

Use representative routine, difficult, and adversarial cases. Repeat tests where behavior varies. Compare with both the current process and a simpler alternative.

For the business case:

Net time released=Baseline human effort−Remaining human effort

Including review and rework.

Net financial benefit=Attributable financial benefit−Implementation and operating costs

Over the same period.

Released capacity is not automatically cash savings. Explain whether it enables more throughput, reduces overtime, avoids hiring, or produces another measurable result. Avoid counting the same benefit twice.

Track severe failures separately from averages. An unauthorized action should trigger investigation even when overall completion rates look strong. No observed incident in a small pilot proves universal safety.

Scale only when the solution meets agreed quality, control, user-value, and cost thresholds. Simplify, redesign, or stop when it does not.

Build, buy, or partner

Should you build, buy, or work with an implementation partner?

Sourcing spectrum

  1. 01Off-the-shelf agent
  2. 02Agent development platform
  3. 03Custom development

Implementation partnerOptions can be combined: an implementation partner can support any point on this spectrum.

Scroll horizontally to compare all columns.

Should you build, buy, or work with an implementation partner?
ApproachMakes sense whenMain trade-off
Off-the-shelf agentA standard workflow fits an existing productFaster setup, with limits on behavior and customization
Agent development platformYour team needs configurable tools, connectors, and deployment capabilitiesLess infrastructure work, with platform dependencies and fees
Custom developmentImportant requirements exceed packaged capabilities and your team can maintain the solutionGreater control, with more engineering responsibility
Implementation partnerYou need delivery expertise or capacity across discovery, integration, evaluation, and rolloutResponsibilities, handover, and support must be contractually clear

These options can be combined. A partner may configure a platform; an internal team may build custom components around purchased services.

Ask every provider to demonstrate your workflow and difficult cases. Compare integration depth, authorization, evaluation, support, and total operating cost.

What determines implementation cost?

Budget for three categories:

  • BuildDiscovery, process design, data preparation, integrations, testing, and user training.
  • RunModel and platform usage, hosting, retrieval, monitoring, human review, and support.
  • ChangeNew exceptions, updated source material, model changes, connector maintenance, and regression testing.

A useful estimate needs scope, transaction volume, system dependencies, quality requirements, and oversight requirements. Public price ranges cannot establish your project budget.

Reduce avoidable lock-in by clarifying ownership, export rights, platform dependencies, and access to instructions, evaluations, logs, and integrations. Switching providers may still require adaptation and retesting.

How Golabs helps

How Golabs helps businesses start with AI agents

Golabs offers custom AI engineering and implementation support for organizations that need help connecting a business problem to a working solution.

Its services address different needs:

  • Tailored AI Agents covers custom agents designed around business workflows, existing systems, and human escalation.
  • AI-Orchestrator provides a dedicated human AI expert to assess opportunities and implement initiatives.
  • AI Fluency Program supports understanding and practical skills across leaders and teams.
  • AI Dedicated Teams provides AI professionals to extend implementation capacity.

The right conversation starts with the workflow: what should improve, what information is available, what actions are permitted, and what evidence would justify further investment.

A packaged solution or internal build may be sufficient. External support becomes useful when expertise, integration complexity, or delivery capacity creates a gap.

FAQs

Frequently asked questions

What should our first AI agent do?

Choose one bounded task with measurable outcomes and manageable consequences. Investigating and preparing support cases for human review is one possible starting point. Validate that it needs adaptive tool use.

Do we need to train our own model?

Not necessarily. First evaluate whether an existing model with appropriate information access and tools meets your requirements. Custom model work needs a specific justification.

Can we start without perfect data?

Yes, if the selected workflow has sufficiently usable information and known limitations. Missing, inaccessible, or unreliable critical information must be addressed or excluded from scope.

Who owns monitoring after the pilot?

Assign ownership before deployment. The business owner tracks outcomes; technical owners manage reliability, access, integrations, and changes. A partner’s responsibilities depend on the agreed engagement.

What happens after a successful pilot?

Expand gradually, continuing evaluation and monitoring. Changes to data, models, and usage can affect performance after deployment. AWS guidance on ongoing evaluation

Your next step

Choose a practical first step

Shortlist three workflows. Name an owner, establish a baseline, and identify the smallest solution worth testing.

If you need help evaluating the use case or its implementation feasibility, bring that workflow to Golabs.

Talk to Golabs About Your First AI Agent